Optimization Of Asphalt Concrete Mix Design to Control Cracking and Rutting Potential
Advisor: Professor Imad L. Al-Qadi
Asphalt concrete (AC) balanced mix design (BMD) approaches are gaining momentum for adoption by state agencies. BMD is based upon the selection of the appropriate aggregate type and gradation, volumetrics, and binder type and content that could control cracking and rutting potential. Thus, the use of AC performance tests, such as Illinois Flexibility Index Test (I-FIT) for cracking potential prediction and Hamburg Wheel Tracking Test (HWTT) for rutting potential prediction, respectively, is expected to increase. The BMD relies on trial-and-error process to optimize I-FIT and HWTT results. Minimizing or eliminating the trial-and-error process would increase productivity and allows AC mix designers to better predict I-FIT and HWTT test results based on AC mix parameters.
In this research an AC mix design framework was developed to complement Superpave AC mix procedures. A deep neural network (DNN) model that relates the impact of constituent materials and AC mixture properties on cracking and rutting potential was introduced. Cracking and rutting potential were measured by the I-FIT and HWTT. The most suitable algorithms, given the data constraints, were applied.
A database of I-FIT and HWTT results was compiled. A total of 18,594 I-FIT datasets were collected from 2061 AC mix designs. For HWTT, 8,263 datasets were collected from 3,782 AC mix designs. A total of 25 input features were identified which included: AC mix type, compaction effort (number of gyration), binder grade, volumetrics properties, aggregate gradation, and reclaimed asphalt pavement content. Data exploration analysis was conducted to evaluate the impact of the AC properties on the I-FIT and the HWTT results. The HWTT rut depth threshold was selected at 5000 passes, while the I-FIT’s flexibility index (FI) was selected at 8. Most of the inputs were found to have an influence on the I-FIT and HWTT results.
Two deep learning models to predict FI and rut depth were developed. Two DNN were then trained to predict FI and rut depth. Monte Carlo Dropout simulations were used in the DNN model to compute a distribution of predicted FI and rut depth. The distribution provides best estimate and range of FI and rut depth. The Monte Carlo Dropout simulations developed a distribution of FI and rut depth predictions with a coefficient of variation (CoV) within 10 to 30%.
Finally, an autoencoder deep neural network (ADNN) was developed to design optimized AC mixes that can meet a prescribed FI and rut depth. The autoencoders are a type of neural network designed for representation learning composed of an encoder and decoder. An autoencoder was trained to predict the AC binder content, and aggregate gradation based on a target mix type, FI, and rut depth. The proposed autoencoder is composed of an encoder of five hidden layers, a latent space of one node and a five hidden layer decoder. The developed model can provide acceptable predictions of aggregate gradation and binder content to meet the required FI and rut depth.